NVIDIA Fermi: Architecture discussion

[DegustatoR]

If Nvidia can't hit their clocks, they may have driver issues and be slower than the competition.

Even at 1260 MHz (which i'm quite sceptical of because there is no mention anywhere that 630 GFlops DP is a theoretical peak of Tesla 20) it won't be slower than Cypress and it will be better against it than GT200/b ever was against RV770/790. Consider this please.

 

[Bouncing Zabaglione Bros.]

Even at 1260 MHz (which i'm quite sceptical of because there is no mention anywhere that 630 GFlops DP is a theoretical peak of Tesla 20) it won't be slower than Cypress and it will be better against it than GT200/b ever was against RV770/790. Consider this please.

If you're going by rumours, then Fermi's TFlops are low. Compared with the cost and power requirements of even one 5870 (let alone several of them in quadfire), then Fermi looks too expensive for what it does, or not enough for what it costs. And it's so far off, AMD might change the landscape again by the time Nvidia get something out there.

I know I'd rather bet on something that exists than promises of something that doesn't.

 

[DegustatoR]

If you're going by rumours, then Fermi's TFlops are low.

TFlops means almost nothing.

 

[Silus]

If you're going by rumours, then Fermi's TFlops are low. Compared with the cost and power requirements of even one 5870 (let alone several of them in quadfire), then Fermi looks too expensive for what it does, or not enough for what it costs. And it's so far off, AMD might change the landscape again by the time Nvidia get something out there.

I know I'd rather bet on something that exists than promises of something that doesn't.

AMD has made a fine job in making you (and many others) believe that "TeraFLOPS" are that important

Higher FLOPS capacity isn't really going to make a card faster than another just by itself and NVIDIA already showed that in the past. Their GTX 280 had half the FLOPS capacity for Single Precision than the HD 4870, yet the GTX 280 was faster. ATI itself also proved that, when they launched the HD 5870 which is not much more than a single chip HD 4870 X2 with DX11 support. Despite more than doubling the peak FLOPs capacity, the HD 5870 is only around 50-60% faster than the HD 4870.

 

[leoneazzurro]

Who's faster depends too much on the workload and specific optimizations... and it's pointless to make comparison in GPGPU based on the gaming comparison and comparison in Gaming using DP FLOPs.
In general, Nvidia has stil the edge in GPGPU, especially because of the Software support that today is better than ATI's in this regard.

 

[Malo]

I highly doubt Nvidia would be touting a card that costs several thousand dollars over something like the 5870 for the HPC market if it didn't have a lot more available to it specifically for that market. This includes CUDA, drivers, specific HPC chip features that aren't required/useful for gaming and probably more features that many here far more knowledgeable could rant on about.

The fact is we really have little information about what the Geforce version of Fermi will look like, and how it will perform. AMD are in a fantastic position at the moment for the desktop market and they should do everything they can to get an edge during this period. I doubt that Nvidia will just bend over and take it though. I really don't think anything can make any solid conclusions about Nvidia's desktop part yet, nor their direction as a company until we have something other than paper.

 

[wishiknew]

Why are people keying in on the flops number. What happened to Z, filtering and fill rates etc ...

 

[Tchock]

If Nvidia can't hit their clocks, they may have driver issues and be slower than the competition. It would be ironic if Nvidia kickstarts the HPC/GPU market and AMD comes in and takes it all away with significantly cheaper and more powerful hardware.

Well they're trying to do that for the professional market with CAD at least (VBO + massive app optimization for the FirePros), but for DCC nVidia's still the preferred choice- I mean they basically put their hands into whatever event CGSociety holds, much better mindshare I guess.

AMD has made a fine job in making you (and many others) believe that "TeraFLOPS" are that important

Higher FLOPS capacity isn't really going to make a card faster than another just by itself and NVIDIA already showed that in the past. Their GTX 280 had half the FLOPS capacity for Single Precision than the HD 4870, yet the GTX 280 was faster. ATI itself also proved that, when they launched the HD 5870 which is not much more than a single chip HD 4870 X2 with DX11 support. Despite more than doubling the peak FLOPs capacity, the HD 5870 is only around 50-60% faster than the HD 4870.

I don't think this post warrants anything more than rolled eyeballs.
So there you go:

 

[trinibwoy]

Why are people keying in on the flops number. What happened to Z, filtering and fill rates etc ...

Cause that's the only number they can peg to Fermi so far. It's amusing though considering that flops were the worst indicator of performance last generation compared to bandwidth, fillrate etc.

 

[Acert93]

AMD has made a fine job in making you (and many others) believe that "TeraFLOPS" are that important

I wouldn't blame AMD singularly for this. NV and others have been just as guilty of playing the FLOPs game.

Why are people keying in on the flops number. What happened to Z, filtering and fill rates etc ...

It all depends on workload. For the workloads they are advertising for FERMI thus far FLOPs and Bandwidth are going to be important, but the architecture will be huge. The problem is we don't have realworld hardware to do practical tests--the problem with paper launches. Live and Die by the same sword I guess.

 

[Malo]

Cause that's the only number they can peg to Fermi so far. It's amusing though considering that flops were the worst indicator of performance last generation compared to bandwidth, fillrate etc.

It's also all the public has to go on since there is no product to actually base numbers for other aspects of rendering. This is pre-launch speculation after all, so there are many speculating as to its performance given limited information. Not really useful yet but isn't it what we like to do, especially so close to release?

 

[Bouncing Zabaglione Bros.]

AMD has made a fine job in making you (and many others) believe that "TeraFLOPS" are that important

And Nvidia does a fine job in making you (and many others) believe that nothing they don't have a lead in is that important. Unfortunately, when Nvidia are so far behind in almost everything bar PhysX, that doesn't leave them a lot to talk about except paper launches of a product that won't be around for another half a year. That's assuming they can eventually make it, and then make a version that's cheap enough to sell to gamers.

What next, Nvidia tells us that getting a product out that works isn't that important? It would go along with the other "DX11/Win7/Christmas isn't that important" statements they've made.

 

[Silus]

Cause that's the only number they can peg to Fermi so far. It's amusing though considering that flops were the worst indicator of performance last generation compared to bandwidth, fillrate etc.

Not just last generation. Ever since the 8800 GTX/HD 2900 XT debuted, that we've witnessed that the card with lower theoretical peak FLOPS capacity, was usually faster in real-world applications.

But, as someone else already mentioned, it obviously depends on the application using it and the type of workload involved.

 

[trinibwoy]

It's also all the public has to go on since there is no product to actually base numbers for other aspects of rendering. This is pre-launch speculation after all, so there are many speculating as to its performance given limited information. Not really useful yet but isn't it what we like to do, especially so close to release?

Yep, that in itself is fine. It's when people attempt to base conclusions around it is when things get iffy.

BZB: jabs at Nvidia aside, where are the real-world indicators of flops being top dog even in HPC? Over in the CUDA forums people are screaming about bandwidth limitations, not flops. It seems that getting data to the shaders is as important as crunching the data when it gets there.

 

[Silus]

And Nvidia does a fine job in making you (and many others) believe that nothing they don't have a lead in is that important. Unfortunately, when Nvidia are so far behind in almost everything bar PhysX, that doesn't leave them a lot to talk about except paper launches of a product that won't be around for another half a year. That's assuming they can eventually make it, and then make a version that's cheap enough to sell to gamers.

Didn't you just describe any company that launches a product later than their competition ?...

As for FLOPs, what I mentioned isn't anything that NVIDIA said. They are facts collected from the usage of the cards in real-world scenarios, not theoretical "mumbo jumbo". If you want to use it as the sole indicator of real performance, be my guest, but that doesn't make it true

 

[Dave Baumann]

Cause that's the only number they can peg to Fermi so far. It's amusing though considering that flops were the worst indicator of performance last generation compared to bandwidth, fillrate etc.

Respective generational architectural progressions though, relative FLOPs can be read and extrapolated.

 

[neliz]

The only benchmarks we've seen so far that use all those FLOPS seriously is Voxilla's benchmarks where a 5870 is four times as fast as a GTX280 or a HD4870 but still not reaching over 1.6TFlop/s

http://forum.beyond3d.com/showthread.php?t=55330

 

[Silus]

I wouldn't blame AMD singularly for this. NV and others have been just as guilty of playing the FLOPs game.

Of course, but I only remember AMD promoting their HD 48xx series with the "over 1 TeraFLOP of power" thingy, as if it mattered. Obviously for Regular Joe, that doesn't even know what that is, it may caused some appeal

 

[trinibwoy]

Respective generational architectural progressions though, relative FLOPs can be read and extrapolated.

Sure, for relatively similiar architectures like RV770 and RV870 that comparison is trivial. But what are you going to extrapolate to with Fermi? There's no MUL, so the contribution of the SFU doesn't even enter into the marketing flop number. The scheduler and memory subsystem are significantly changed and there's the addition of predication for all instructions etc. Extrapolaton doesn't look all that straightforward to me.

Of course, but I only remember AMD promoting their HD 48xx series with the "over 1 TeraFLOP of power" thingy, as if it mattered. Obviously for Regular Joe, that doesn't even know what that is, it may caused some appeal

Nvidia would've too if they had hit 1 teraflop

 

[neliz]

Of course, but I only remember AMD promoting their HD 48xx series with the "over 1 TeraFLOP of power" thingy, as if it mattered. Obviously for Regular Joe, that doesn't even know what that is, it may caused some appeal

It was Michael Hara that started it before G92 came out, he already mentioned that GT200 would be close to 1TFLOP/s (May 2007.) Later they promoted that their Tesla board would get between 1 and 1.33 TFlOP/s. they just got 1-UPed (remember the moon landing picture.) This was all posted by Arun.

http://www.howtofixcomputers.com/fo...rformance-1-1-33-tflop-155136.html#post633554